Dear Yan,
In my opinion, the best option of the analysis is the last one (p<0.001 uncorrected + SVC). This is a widely acceptable approach and, if you have a strong hypothesis concerning amygdala, no one will criticize you.
Here I try to propose my understanding of your cluster vs voxel results. I am not a statistician, so, take my opinion carefully.
In all of your analyses the SVC significance of the peak voxel [-24 -4 -18] is the same, and it is not surprising, since T statics of this particular voxel did not vary across the threshold levels you used.
As for the cluster level significance, here we see that the threshold used has an impact. The size of the cluster will be obviousely greater for higher p threshold. As an extreme case,
if you choose to set p=1 the sluster size will be equal (or close) to the size of your volume. But the criterion of the cluster-level significance also changes as a function of the threshold. You can see this at the end of each table: expected foxel per cluster (for the brain volume) is 28 for p<0.001, 51 for p<0.005 and 247 for p<0.05. Of course, if you use SVC, this criterion will be much less severe, since the whole volume of your are is much smaller that the brain volume, thus, the likelihood of a cluster of 3 contiguous voxels is already low enough, to call this a significant cluster.
Hope my input can be helpful.
Sincerely yours,
Vladimir
Volodymyr B. Bogdanov, PhD
Nutrition and Integrative Neurobiology (NutriNeuro)
UMR INRA 1286 - Université Bordeaux 2
UFR Pharmacie 2e Tranche 2e étage
Case courrier 34
146 rue Léo Saignat
33076 BORDEAUX Cedex
FRANCE
Dear all,
I was wondering what was the ideal
way to report SVC results: at the cluster or the voxel level?
I attached a screenshot (called SVC1) of my analysis (two-sample t-test) showing both significant result at the voxel and at the cluster level. I used in that case an uncorrected p=.005 and then searched in my ROI (ie left amygdala).
Surprisingly, I had a look at a more liberal threshold (p=.05), and only the voxel level was significant (screenshot called SVC2). Also, I tried it at p=.001, and both cluster and voxel levels were significant, with a cluster size of 3 voxels only (screenshot SVC3)!
I am not sure I understand this result. Any thought? What option will be the more powerful and/or acceptable? Or maybe I am completely wrong?
Thanks for the tips!
Cheers,
Yann
